* Havard Eidnes
If I've understood correctly, the "P" in "PA" (and "PI") is meant to be more or less synonymous with ISP, not with a provider of LIR services only. This was so that the ISP could announce the whole covering address space as a single route, thereby reducing the amount of entropy we collectively have to carry on our backs as ISPs. If the ISP / PA block holder insists, and you as a customer and current sub-PA-block holder wish to cancel the service with the ISP, the ISP can insist that you cease using the PA addresses you were assigned as a customer.
The converse is not true: if the PA-holding LIR lets you take your sub-block with you, they can allow it, and I beleive that's what you said as well, Tore. I'm not sure that is the typical case, though(?)
It's not the common case, no. Usually, ISP = LIR. But even though it might be a corner case, it's still a case we do (and should) cater for.
Your example with a government or large organization which holds one or more large PA block and which out of administrative convenience ("renumbering is so hard, even if I just have client hosts!") or for other reasons doles out address blocks to widely distributed sub- organizations, and where each sub-organization is free to choose its own ISP to use will result in injection of more entropy into the global routing system, as each individual sub-organization's route will need to be carried globally, and there's no possibility for route aggregation. I'm hesitating a little to find an appropriate characterization of what would happen if such pratices became very widespread, but I'm sure it certainly isn't positive for the sustainability of the network.
Regretfully, noone has come up with any sort of economic (the only one which works...) dis-incentive countering such behaviour, so we'll end up by muddling along.
BTW, this argument is address-family independent...
Indeed. But, the reason why such non-ISP LIRs might become more prevalent nowadays is IPv4 depletion. We already know that IPv4 isn't going to be sustainable though, so I don't think it is anything we need to worry about or attempt to "fix" or "prevent" through implementing restrictive policies. In the longer term, in the IPv6 world, such non-ISP LIRs will again be just a corner case that exists in limited numbers, and probably only where there's actually a good reason for doing it that way. Allowing for them to exist thus won't cause significant harm to the routing table. (Also keep in mind that the preferred alternative for many of them would be to use IPv6 PI instead, which would be worse as a bunch of PI blocks cannot be re-aggregated at a later point in time.) Tore